ctdb/server/ctdb_lock.c

   1 /*
   2    ctdb lock handling
   3    provide API to do non-blocking locks for single or all databases
   4
   5    Copyright (C) Amitay Isaacs  2012
   6
   7    This program is free software; you can redistribute it and/or modify
   8    it under the terms of the GNU General Public License as published by
   9    the Free Software Foundation; either version 3 of the License, or
  10    (at your option) any later version.
  11
  12    This program is distributed in the hope that it will be useful,
  13    but WITHOUT ANY WARRANTY; without even the implied warranty of
  14    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15    GNU General Public License for more details.
  16
  17    You should have received a copy of the GNU General Public License
  18    along with this program; if not, see <http://www.gnu.org/licenses/>.
  19 */
  20 #include "includes.h"
  21 #include "include/ctdb_private.h"
  22 #include "include/ctdb_protocol.h"
  23 #include "tevent.h"
  24 #include "tdb.h"
  25 #include "lib/tdb_wrap/tdb_wrap.h"
  26 #include "system/filesys.h"
  27 #include "lib/util/dlinklist.h"
  28
  29 /*
  30  * Non-blocking Locking API
  31  *
  32  * 1. Create a child process to do blocking locks.
  33  * 2. Once the locks are obtained, signal parent process via fd.
  34  * 3. Invoke registered callback routine with locking status.
  35  * 4. If the child process cannot get locks within certain time,
  36  *    execute an external script to debug.
  37  *
  38  * ctdb_lock_record()      - get a lock on a record
  39  * ctdb_lock_db()          - get a lock on a DB
  40  * ctdb_lock_alldb_prio()  - get a lock on all DBs with given priority
  41  * ctdb_lock_alldb()       - get a lock on all DBs
  42  *
  43  *  auto_mark              - whether to mark/unmark DBs in before/after callback
  44  */
  45
  46 enum lock_type {
  47         LOCK_RECORD,
  48         LOCK_DB,
  49         LOCK_ALLDB_PRIO,
  50         LOCK_ALLDB,
  51 };
  52
  53 static const char * const lock_type_str[] = {
  54         "lock_record",
  55         "lock_db",
  56         "lock_alldb_prio",
  57         "lock_alldb",
  58 };
  59
  60 struct lock_request;
  61
  62 /* lock_context is the common part for a lock request */
  63 struct lock_context {
  64         struct lock_context *next, *prev;
  65         enum lock_type type;
  66         struct ctdb_context *ctdb;
  67         struct ctdb_db_context *ctdb_db;
  68         TDB_DATA key;
  69         uint32_t priority;
  70         bool auto_mark;
  71         struct lock_request *request;
  72         pid_t child;
  73         int fd[2];
  74         struct tevent_fd *tfd;
  75         struct tevent_timer *ttimer;
  76         struct timeval start_time;
  77         uint32_t key_hash;
  78         bool can_schedule;
  79 };
  80
  81 /* lock_request is the client specific part for a lock request */
  82 struct lock_request {
  83         struct lock_context *lctx;
  84         void (*callback)(void *, bool);
  85         void *private_data;
  86 };
  87
  88
  89 /*
  90  * Support samba 3.6.x (and older) versions which do not set db priority.
  91  *
  92  * By default, all databases are set to priority 1. So only when priority
  93  * is set to 1, check for databases that need higher priority.
  94  */
  95 static bool later_db(struct ctdb_context *ctdb, const char *name)
  96 {
  97         if (ctdb->tunable.samba3_hack == 0) {
  98                 return false;
  99         }
 100
 101         if (strstr(name, "brlock") ||
 102             strstr(name, "g_lock") ||
 103             strstr(name, "notify_onelevel") ||
 104             strstr(name, "serverid") ||
 105             strstr(name, "xattr_tdb")) {
 106                 return true;
 107         }
 108
 109         return false;
 110 }
 111
 112 typedef int (*db_handler_t)(struct ctdb_db_context *ctdb_db,
 113                             uint32_t priority,
 114                             void *private_data);
 115
 116 static int ctdb_db_iterator(struct ctdb_context *ctdb, uint32_t priority,
 117                             db_handler_t handler, void *private_data)
 118 {
 119         struct ctdb_db_context *ctdb_db;
 120         int ret;
 121
 122         for (ctdb_db = ctdb->db_list; ctdb_db; ctdb_db = ctdb_db->next) {
 123                 if (ctdb_db->priority != priority) {
 124                         continue;
 125                 }
 126                 if (later_db(ctdb, ctdb_db->db_name)) {
 127                         continue;
 128                 }
 129                 ret = handler(ctdb_db, priority, private_data);
 130                 if (ret != 0) {
 131                         return -1;
 132                 }
 133         }
 134
 135         /* If priority != 1, later_db check is not required and can return */
 136         if (priority != 1) {
 137                 return 0;
 138         }
 139
 140         for (ctdb_db = ctdb->db_list; ctdb_db; ctdb_db = ctdb_db->next) {
 141                 if (!later_db(ctdb, ctdb_db->db_name)) {
 142                         continue;
 143                 }
 144                 ret = handler(ctdb_db, priority, private_data);
 145                 if (ret != 0) {
 146                         return -1;
 147                 }
 148         }
 149
 150         return 0;
 151 }
 152
 153
 154 /*
 155  * lock all databases - mark only
 156  */
 157 static int db_lock_mark_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 158                                 void *private_data)
 159 {
 160         int tdb_transaction_write_lock_mark(struct tdb_context *);
 161
 162         DEBUG(DEBUG_INFO, ("marking locked database %s, priority:%u\n",
 163                            ctdb_db->db_name, priority));
 164
 165         if (tdb_transaction_write_lock_mark(ctdb_db->ltdb->tdb) != 0) {
 166                 DEBUG(DEBUG_ERR, ("Failed to mark (transaction lock) database %s\n",
 167                                   ctdb_db->db_name));
 168                 return -1;
 169         }
 170
 171         if (tdb_lockall_mark(ctdb_db->ltdb->tdb) != 0) {
 172                 DEBUG(DEBUG_ERR, ("Failed to mark (all lock) database %s\n",
 173                                   ctdb_db->db_name));
 174                 return -1;
 175         }
 176
 177         return 0;
 178 }
 179
 180 int ctdb_lockall_mark_prio(struct ctdb_context *ctdb, uint32_t priority)
 181 {
 182         /*
 183          * This function is only used by the main dameon during recovery.
 184          * At this stage, the databases have already been locked, by a
 185          * dedicated child process. The freeze_mode variable is used to track
 186          * whether the actual locks are held by the child process or not.
 187          */
 188
 189         if (ctdb->freeze_mode[priority] != CTDB_FREEZE_FROZEN) {
 190                 DEBUG(DEBUG_ERR, ("Attempt to mark all databases locked when not frozen\n"));
 191                 return -1;
 192         }
 193
 194         return ctdb_db_iterator(ctdb, priority, db_lock_mark_handler, NULL);
 195 }
 196
 197 static int ctdb_lockall_mark(struct ctdb_context *ctdb)
 198 {
 199         uint32_t priority;
 200
 201         for (priority=1; priority<=NUM_DB_PRIORITIES; priority++) {
 202                 if (ctdb_db_iterator(ctdb, priority, db_lock_mark_handler, NULL) != 0) {
 203                         return -1;
 204                 }
 205         }
 206
 207         return 0;
 208 }
 209
 210
 211 /*
 212  * lock all databases - unmark only
 213  */
 214 static int db_lock_unmark_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 215                                   void *private_data)
 216 {
 217         int tdb_transaction_write_lock_unmark(struct tdb_context *);
 218
 219         DEBUG(DEBUG_INFO, ("unmarking locked database %s, priority:%u\n",
 220                            ctdb_db->db_name, priority));
 221
 222         if (tdb_transaction_write_lock_unmark(ctdb_db->ltdb->tdb) != 0) {
 223                 DEBUG(DEBUG_ERR, ("Failed to unmark (transaction lock) database %s\n",
 224                                   ctdb_db->db_name));
 225                 return -1;
 226         }
 227
 228         if (tdb_lockall_unmark(ctdb_db->ltdb->tdb) != 0) {
 229                 DEBUG(DEBUG_ERR, ("Failed to unmark (all lock) database %s\n",
 230                                   ctdb_db->db_name));
 231                 return -1;
 232         }
 233
 234         return 0;
 235 }
 236
 237 int ctdb_lockall_unmark_prio(struct ctdb_context *ctdb, uint32_t priority)
 238 {
 239         /*
 240          * This function is only used by the main daemon during recovery.
 241          * At this stage, the databases have already been locked, by a
 242          * dedicated child process. The freeze_mode variable is used to track
 243          * whether the actual locks are held by the child process or not.
 244          */
 245
 246         if (ctdb->freeze_mode[priority] != CTDB_FREEZE_FROZEN) {
 247                 DEBUG(DEBUG_ERR, ("Attempt to unmark all databases locked when not frozen\n"));
 248                 return -1;
 249         }
 250
 251         return ctdb_db_iterator(ctdb, priority, db_lock_unmark_handler, NULL);
 252 }
 253
 254 static int ctdb_lockall_unmark(struct ctdb_context *ctdb)
 255 {
 256         uint32_t priority;
 257
 258         for (priority=NUM_DB_PRIORITIES; priority>0; priority--) {
 259                 if (ctdb_db_iterator(ctdb, priority, db_lock_unmark_handler, NULL) != 0) {
 260                         return -1;
 261                 }
 262         }
 263
 264         return 0;
 265 }
 266
 267
 268 static void ctdb_lock_schedule(struct ctdb_context *ctdb);
 269
 270 /*
 271  * Destructor to kill the child locking process
 272  */
 273 static int ctdb_lock_context_destructor(struct lock_context *lock_ctx)
 274 {
 275         if (lock_ctx->request) {
 276                 lock_ctx->request->lctx = NULL;
 277         }
 278         if (lock_ctx->child > 0) {
 279                 ctdb_kill(lock_ctx->ctdb, lock_ctx->child, SIGKILL);
 280                 if (lock_ctx->type == LOCK_RECORD) {
 281                         DLIST_REMOVE(lock_ctx->ctdb_db->lock_current, lock_ctx);
 282                 } else {
 283                         DLIST_REMOVE(lock_ctx->ctdb->lock_current, lock_ctx);
 284                 }
 285                 if (lock_ctx->ctdb_db) {
 286                         lock_ctx->ctdb_db->lock_num_current--;
 287                 }
 288                 CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_current);
 289                 if (lock_ctx->ctdb_db) {
 290                         CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_current);
 291                 }
 292         } else {
 293                 if (lock_ctx->type == LOCK_RECORD) {
 294                         DLIST_REMOVE(lock_ctx->ctdb_db->lock_pending, lock_ctx);
 295                 } else {
 296                         DLIST_REMOVE(lock_ctx->ctdb->lock_pending, lock_ctx);
 297                 }
 298                 CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_pending);
 299                 if (lock_ctx->ctdb_db) {
 300                         CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
 301                 }
 302         }
 303
 304         ctdb_lock_schedule(lock_ctx->ctdb);
 305
 306         return 0;
 307 }
 308
 309
 310 /*
 311  * Destructor to remove lock request
 312  */
 313 static int ctdb_lock_request_destructor(struct lock_request *lock_request)
 314 {
 315         TALLOC_FREE(lock_request->lctx);
 316         return 0;
 317 }
 318
 319 /*
 320  * Process all the callbacks waiting for lock
 321  *
 322  * If lock has failed, callback is executed with locked=false
 323  */
 324 static void process_callbacks(struct lock_context *lock_ctx, bool locked)
 325 {
 326         struct lock_request *request;
 327
 328         if (lock_ctx->auto_mark && locked) {
 329                 switch (lock_ctx->type) {
 330                 case LOCK_RECORD:
 331                         tdb_chainlock_mark(lock_ctx->ctdb_db->ltdb->tdb, lock_ctx->key);
 332                         break;
 333
 334                 case LOCK_DB:
 335                         tdb_lockall_mark(lock_ctx->ctdb_db->ltdb->tdb);
 336                         break;
 337
 338                 case LOCK_ALLDB_PRIO:
 339                         ctdb_lockall_mark_prio(lock_ctx->ctdb, lock_ctx->priority);
 340                         break;
 341
 342                 case LOCK_ALLDB:
 343                         ctdb_lockall_mark(lock_ctx->ctdb);
 344                         break;
 345                 }
 346         }
 347
 348         request = lock_ctx->request;
 349         if (lock_ctx->auto_mark) {
 350                 /* Reset the destructor, so request is not removed from the list */
 351                 talloc_set_destructor(request, NULL);
 352         }
 353         request->callback(request->private_data, locked);
 354
 355         if (lock_ctx->auto_mark && locked) {
 356                 switch (lock_ctx->type) {
 357                 case LOCK_RECORD:
 358                         tdb_chainlock_unmark(lock_ctx->ctdb_db->ltdb->tdb, lock_ctx->key);
 359                         break;
 360
 361                 case LOCK_DB:
 362                         tdb_lockall_unmark(lock_ctx->ctdb_db->ltdb->tdb);
 363                         break;
 364
 365                 case LOCK_ALLDB_PRIO:
 366                         ctdb_lockall_unmark_prio(lock_ctx->ctdb, lock_ctx->priority);
 367                         break;
 368
 369                 case LOCK_ALLDB:
 370                         ctdb_lockall_unmark(lock_ctx->ctdb);
 371                         break;
 372                 }
 373         }
 374 }
 375
 376
 377 static int lock_bucket_id(double t)
 378 {
 379         double ms = 1.e-3, s = 1;
 380         int id;
 381
 382         if (t < 1*ms) {
 383                 id = 0;
 384         } else if (t < 10*ms) {
 385                 id = 1;
 386         } else if (t < 100*ms) {
 387                 id = 2;
 388         } else if (t < 1*s) {
 389                 id = 3;
 390         } else if (t < 2*s) {
 391                 id = 4;
 392         } else if (t < 4*s) {
 393                 id = 5;
 394         } else if (t < 8*s) {
 395                 id = 6;
 396         } else if (t < 16*s) {
 397                 id = 7;
 398         } else if (t < 32*s) {
 399                 id = 8;
 400         } else if (t < 64*s) {
 401                 id = 9;
 402         } else {
 403                 id = 10;
 404         }
 405
 406         return id;
 407 }
 408
 409 /*
 410  * Callback routine when the required locks are obtained.
 411  * Called from parent context
 412  */
 413 static void ctdb_lock_handler(struct tevent_context *ev,
 414                             struct tevent_fd *tfd,
 415                             uint16_t flags,
 416                             void *private_data)
 417 {
 418         struct lock_context *lock_ctx;
 419         TALLOC_CTX *tmp_ctx = NULL;
 420         char c;
 421         bool locked;
 422         double t;
 423         int id;
 424
 425         lock_ctx = talloc_get_type_abort(private_data, struct lock_context);
 426
 427         /* cancel the timeout event */
 428         TALLOC_FREE(lock_ctx->ttimer);
 429
 430         t = timeval_elapsed(&lock_ctx->start_time);
 431         id = lock_bucket_id(t);
 432
 433         if (lock_ctx->auto_mark) {
 434                 tmp_ctx = talloc_new(ev);
 435                 talloc_steal(tmp_ctx, lock_ctx);
 436         }
 437
 438         /* Read the status from the child process */
 439         if (sys_read(lock_ctx->fd[0], &c, 1) != 1) {
 440                 locked = false;
 441         } else {
 442                 locked = (c == 0 ? true : false);
 443         }
 444
 445         /* Update statistics */
 446         CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_calls);
 447         if (lock_ctx->ctdb_db) {
 448                 CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_calls);
 449         }
 450
 451         if (locked) {
 452                 if (lock_ctx->ctdb_db) {
 453                         CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.buckets[id]);
 454                         CTDB_UPDATE_LATENCY(lock_ctx->ctdb, lock_ctx->ctdb_db,
 455                                             lock_type_str[lock_ctx->type], locks.latency,
 456                                             lock_ctx->start_time);
 457
 458                         CTDB_UPDATE_DB_LATENCY(lock_ctx->ctdb_db, lock_type_str[lock_ctx->type], locks.latency, t);
 459                         CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.buckets[id]);
 460                 }
 461         } else {
 462                 CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_failed);
 463                 if (lock_ctx->ctdb_db) {
 464                         CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_failed);
 465                 }
 466         }
 467
 468         process_callbacks(lock_ctx, locked);
 469
 470         if (lock_ctx->auto_mark) {
 471                 talloc_free(tmp_ctx);
 472         }
 473 }
 474
 475
 476 /*
 477  * Callback routine when required locks are not obtained within timeout
 478  * Called from parent context
 479  */
 480 static void ctdb_lock_timeout_handler(struct tevent_context *ev,
 481                                     struct tevent_timer *ttimer,
 482                                     struct timeval current_time,
 483                                     void *private_data)
 484 {
 485         static const char * debug_locks = NULL;
 486         struct lock_context *lock_ctx;
 487         struct ctdb_context *ctdb;
 488         pid_t pid;
 489
 490         lock_ctx = talloc_get_type_abort(private_data, struct lock_context);
 491         ctdb = lock_ctx->ctdb;
 492
 493         /* If a node stopped/banned, don't spam the logs */
 494         if (ctdb->nodes[ctdb->pnn]->flags & NODE_FLAGS_INACTIVE) {
 495                 lock_ctx->ttimer = NULL;
 496                 return;
 497         }
 498         if (lock_ctx->ctdb_db) {
 499                 DEBUG(DEBUG_WARNING,
 500                       ("Unable to get %s lock on database %s for %.0lf seconds\n",
 501                        (lock_ctx->type == LOCK_RECORD ? "RECORD" : "DB"),
 502                        lock_ctx->ctdb_db->db_name,
 503                        timeval_elapsed(&lock_ctx->start_time)));
 504         } else {
 505                 DEBUG(DEBUG_WARNING,
 506                       ("Unable to get ALLDB locks for %.0lf seconds\n",
 507                        timeval_elapsed(&lock_ctx->start_time)));
 508         }
 509
 510         /* Fire a child process to find the blocking process. */
 511         if (debug_locks == NULL) {
 512                 debug_locks = getenv("CTDB_DEBUG_LOCKS");
 513                 if (debug_locks == NULL) {
 514                         debug_locks = talloc_asprintf(ctdb,
 515                                                       "%s/debug_locks.sh",
 516                                                       getenv("CTDB_BASE"));
 517                 }
 518         }
 519         if (debug_locks != NULL) {
 520                 pid = vfork();
 521                 if (pid == 0) {
 522                         execl(debug_locks, debug_locks, NULL);
 523                         _exit(0);
 524                 }
 525                 ctdb_track_child(ctdb, pid);
 526         } else {
 527                 DEBUG(DEBUG_WARNING,
 528                       (__location__
 529                        " Unable to setup lock debugging - no memory?\n"));
 530         }
 531
 532         /* reset the timeout timer */
 533         // talloc_free(lock_ctx->ttimer);
 534         lock_ctx->ttimer = tevent_add_timer(ctdb->ev,
 535                                             lock_ctx,
 536                                             timeval_current_ofs(10, 0),
 537                                             ctdb_lock_timeout_handler,
 538                                             (void *)lock_ctx);
 539 }
 540
 541
 542 static int db_count_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 543                             void *private_data)
 544 {
 545         int *count = (int *)private_data;
 546
 547         (*count) += 2;
 548
 549         return 0;
 550 }
 551
 552 static int db_flags(struct ctdb_db_context *ctdb_db)
 553 {
 554         int tdb_flags = TDB_DEFAULT;
 555
 556 #ifdef TDB_MUTEX_LOCKING
 557         if (!ctdb_db->persistent && ctdb_db->ctdb->tunable.mutex_enabled) {
 558                 tdb_flags = (TDB_MUTEX_LOCKING | TDB_CLEAR_IF_FIRST);
 559         }
 560 #endif
 561         return tdb_flags;
 562 }
 563
 564 struct db_namelist {
 565         const char **names;
 566         int n;
 567 };
 568
 569 static int db_name_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 570                            void *private_data)
 571 {
 572         struct db_namelist *list = (struct db_namelist *)private_data;
 573
 574         list->names[list->n] = talloc_strdup(list->names, ctdb_db->db_path);
 575         list->names[list->n+1] = talloc_asprintf(list->names, "0x%x",
 576                                                  db_flags(ctdb_db));
 577         list->n += 2;
 578
 579         return 0;
 580 }
 581
 582 static bool lock_helper_args(TALLOC_CTX *mem_ctx,
 583                              struct lock_context *lock_ctx, int fd,
 584                              int *argc, const char ***argv)
 585 {
 586         struct ctdb_context *ctdb = lock_ctx->ctdb;
 587         const char **args = NULL;
 588         int nargs, i;
 589         int priority;
 590         struct db_namelist list;
 591
 592         switch (lock_ctx->type) {
 593         case LOCK_RECORD:
 594                 nargs = 6;
 595                 break;
 596
 597         case LOCK_DB:
 598                 nargs = 5;
 599                 break;
 600
 601         case LOCK_ALLDB_PRIO:
 602                 nargs = 3;
 603                 ctdb_db_iterator(ctdb, lock_ctx->priority, db_count_handler, &nargs);
 604                 break;
 605
 606         case LOCK_ALLDB:
 607                 nargs = 3;
 608                 for (priority=1; priority<NUM_DB_PRIORITIES; priority++) {
 609                         ctdb_db_iterator(ctdb, priority, db_count_handler, &nargs);
 610                 }
 611                 break;
 612         }
 613
 614         /* Add extra argument for null termination */
 615         nargs++;
 616
 617         args = talloc_array(mem_ctx, const char *, nargs);
 618         if (args == NULL) {
 619                 return false;
 620         }
 621
 622         args[0] = talloc_asprintf(args, "%d", getpid());
 623         args[1] = talloc_asprintf(args, "%d", fd);
 624
 625         switch (lock_ctx->type) {
 626         case LOCK_RECORD:
 627                 args[2] = talloc_strdup(args, "RECORD");
 628                 args[3] = talloc_strdup(args, lock_ctx->ctdb_db->db_path);
 629                 args[4] = talloc_asprintf(args, "0x%x",
 630                                           db_flags(lock_ctx->ctdb_db));
 631                 if (lock_ctx->key.dsize == 0) {
 632                         args[5] = talloc_strdup(args, "NULL");
 633                 } else {
 634                         args[5] = hex_encode_talloc(args, lock_ctx->key.dptr, lock_ctx->key.dsize);
 635                 }
 636                 break;
 637
 638         case LOCK_DB:
 639                 args[2] = talloc_strdup(args, "DB");
 640                 args[3] = talloc_strdup(args, lock_ctx->ctdb_db->db_path);
 641                 args[4] = talloc_asprintf(args, "0x%x",
 642                                           db_flags(lock_ctx->ctdb_db));
 643                 break;
 644
 645         case LOCK_ALLDB_PRIO:
 646                 args[2] = talloc_strdup(args, "DB");
 647                 list.names = args;
 648                 list.n = 3;
 649                 ctdb_db_iterator(ctdb, lock_ctx->priority, db_name_handler, &list);
 650                 break;
 651
 652         case LOCK_ALLDB:
 653                 args[2] = talloc_strdup(args, "DB");
 654                 list.names = args;
 655                 list.n = 3;
 656                 for (priority=1; priority<NUM_DB_PRIORITIES; priority++) {
 657                         ctdb_db_iterator(ctdb, priority, db_name_handler, &list);
 658                 }
 659                 break;
 660         }
 661
 662         /* Make sure last argument is NULL */
 663         args[nargs-1] = NULL;
 664
 665         for (i=0; i<nargs-1; i++) {
 666                 if (args[i] == NULL) {
 667                         talloc_free(args);
 668                         return false;
 669                 }
 670         }
 671
 672         *argc = nargs;
 673         *argv = args;
 674         return true;
 675 }
 676
 677 /*
 678  * Find a lock request that can be scheduled
 679  */
 680 static struct lock_context *ctdb_find_lock_context(struct ctdb_context *ctdb)
 681 {
 682         struct lock_context *lock_ctx, *next_ctx;
 683         struct ctdb_db_context *ctdb_db;
 684
 685         /* First check if there are database lock requests */
 686
 687         for (lock_ctx = ctdb->lock_pending; lock_ctx != NULL;
 688              lock_ctx = next_ctx) {
 689
 690                 if (lock_ctx->request != NULL) {
 691                         /* Found a lock context with a request */
 692                         return lock_ctx;
 693                 }
 694
 695                 next_ctx = lock_ctx->next;
 696
 697                 DEBUG(DEBUG_INFO, ("Removing lock context without lock "
 698                                    "request\n"));
 699                 DLIST_REMOVE(ctdb->lock_pending, lock_ctx);
 700                 CTDB_DECREMENT_STAT(ctdb, locks.num_pending);
 701                 if (lock_ctx->ctdb_db) {
 702                         CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db,
 703                                                locks.num_pending);
 704                 }
 705                 talloc_free(lock_ctx);
 706         }
 707
 708         /* Next check database queues */
 709         for (ctdb_db = ctdb->db_list; ctdb_db; ctdb_db = ctdb_db->next) {
 710                 if (ctdb_db->lock_num_current ==
 711                     ctdb->tunable.lock_processes_per_db) {
 712                         continue;
 713                 }
 714
 715                 for (lock_ctx = ctdb_db->lock_pending; lock_ctx != NULL;
 716                      lock_ctx = next_ctx) {
 717
 718                         next_ctx = lock_ctx->next;
 719
 720                         if (lock_ctx->request != NULL) {
 721                                 return lock_ctx;
 722                         }
 723
 724                         DEBUG(DEBUG_INFO, ("Removing lock context without "
 725                                            "lock request\n"));
 726                         DLIST_REMOVE(ctdb_db->lock_pending, lock_ctx);
 727                         CTDB_DECREMENT_STAT(ctdb, locks.num_pending);
 728                         CTDB_DECREMENT_DB_STAT(ctdb_db, locks.num_pending);
 729                         talloc_free(lock_ctx);
 730                 }
 731         }
 732
 733         return NULL;
 734 }
 735
 736 /*
 737  * Schedule a new lock child process
 738  * Set up callback handler and timeout handler
 739  */
 740 static void ctdb_lock_schedule(struct ctdb_context *ctdb)
 741 {
 742         struct lock_context *lock_ctx;
 743         int ret, argc;
 744         TALLOC_CTX *tmp_ctx;
 745         const char *helper = CTDB_HELPER_BINDIR "/ctdb_lock_helper";
 746         static const char *prog = NULL;
 747         const char **args;
 748
 749         if (prog == NULL) {
 750                 const char *t;
 751
 752                 t = getenv("CTDB_LOCK_HELPER");
 753                 if (t != NULL) {
 754                         prog = talloc_strdup(ctdb, t);
 755                 } else {
 756                         prog = talloc_strdup(ctdb, helper);
 757                 }
 758                 CTDB_NO_MEMORY_VOID(ctdb, prog);
 759         }
 760
 761         /* Find a lock context with requests */
 762         lock_ctx = ctdb_find_lock_context(ctdb);
 763         if (lock_ctx == NULL) {
 764                 return;
 765         }
 766
 767         lock_ctx->child = -1;
 768         ret = pipe(lock_ctx->fd);
 769         if (ret != 0) {
 770                 DEBUG(DEBUG_ERR, ("Failed to create pipe in ctdb_lock_schedule\n"));
 771                 return;
 772         }
 773
 774         set_close_on_exec(lock_ctx->fd[0]);
 775
 776         /* Create data for child process */
 777         tmp_ctx = talloc_new(lock_ctx);
 778         if (tmp_ctx == NULL) {
 779                 DEBUG(DEBUG_ERR, ("Failed to allocate memory for helper args\n"));
 780                 close(lock_ctx->fd[0]);
 781                 close(lock_ctx->fd[1]);
 782                 return;
 783         }
 784
 785         /* Create arguments for lock helper */
 786         if (!lock_helper_args(tmp_ctx, lock_ctx, lock_ctx->fd[1],
 787                               &argc, &args)) {
 788                 DEBUG(DEBUG_ERR, ("Failed to create lock helper args\n"));
 789                 close(lock_ctx->fd[0]);
 790                 close(lock_ctx->fd[1]);
 791                 talloc_free(tmp_ctx);
 792                 return;
 793         }
 794
 795         if (!ctdb_vfork_with_logging(lock_ctx, ctdb, "lock_helper",
 796                                      prog, argc, (const char **)args,
 797                                      NULL, NULL, &lock_ctx->child)) {
 798                 DEBUG(DEBUG_ERR, ("Failed to create a child in ctdb_lock_schedule\n"));
 799                 close(lock_ctx->fd[0]);
 800                 close(lock_ctx->fd[1]);
 801                 talloc_free(tmp_ctx);
 802                 return;
 803         }
 804
 805         /* Parent process */
 806         close(lock_ctx->fd[1]);
 807
 808         talloc_set_destructor(lock_ctx, ctdb_lock_context_destructor);
 809
 810         talloc_free(tmp_ctx);
 811
 812         /* Set up timeout handler */
 813         lock_ctx->ttimer = tevent_add_timer(ctdb->ev,
 814                                             lock_ctx,
 815                                             timeval_current_ofs(10, 0),
 816                                             ctdb_lock_timeout_handler,
 817                                             (void *)lock_ctx);
 818         if (lock_ctx->ttimer == NULL) {
 819                 ctdb_kill(ctdb, lock_ctx->child, SIGKILL);
 820                 lock_ctx->child = -1;
 821                 talloc_set_destructor(lock_ctx, NULL);
 822                 close(lock_ctx->fd[0]);
 823                 return;
 824         }
 825
 826         /* Set up callback */
 827         lock_ctx->tfd = tevent_add_fd(ctdb->ev,
 828                                       lock_ctx,
 829                                       lock_ctx->fd[0],
 830                                       EVENT_FD_READ,
 831                                       ctdb_lock_handler,
 832                                       (void *)lock_ctx);
 833         if (lock_ctx->tfd == NULL) {
 834                 TALLOC_FREE(lock_ctx->ttimer);
 835                 ctdb_kill(ctdb, lock_ctx->child, SIGKILL);
 836                 lock_ctx->child = -1;
 837                 talloc_set_destructor(lock_ctx, NULL);
 838                 close(lock_ctx->fd[0]);
 839                 return;
 840         }
 841         tevent_fd_set_auto_close(lock_ctx->tfd);
 842
 843         /* Move the context from pending to current */
 844         if (lock_ctx->type == LOCK_RECORD) {
 845                 DLIST_REMOVE(lock_ctx->ctdb_db->lock_pending, lock_ctx);
 846                 DLIST_ADD_END(lock_ctx->ctdb_db->lock_current, lock_ctx, NULL);
 847         } else {
 848                 DLIST_REMOVE(ctdb->lock_pending, lock_ctx);
 849                 DLIST_ADD_END(ctdb->lock_current, lock_ctx, NULL);
 850         }
 851         CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_pending);
 852         CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_current);
 853         if (lock_ctx->ctdb_db) {
 854                 lock_ctx->ctdb_db->lock_num_current++;
 855                 CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
 856                 CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_current);
 857         }
 858 }
 859
 860
 861 /*
 862  * Lock record / db depending on type
 863  */
 864 static struct lock_request *ctdb_lock_internal(TALLOC_CTX *mem_ctx,
 865                                                struct ctdb_context *ctdb,
 866                                                struct ctdb_db_context *ctdb_db,
 867                                                TDB_DATA key,
 868                                                uint32_t priority,
 869                                                void (*callback)(void *, bool),
 870                                                void *private_data,
 871                                                enum lock_type type,
 872                                                bool auto_mark)
 873 {
 874         struct lock_context *lock_ctx = NULL;
 875         struct lock_request *request;
 876
 877         if (callback == NULL) {
 878                 DEBUG(DEBUG_WARNING, ("No callback function specified, not locking\n"));
 879                 return NULL;
 880         }
 881
 882         lock_ctx = talloc_zero(ctdb, struct lock_context);
 883         if (lock_ctx == NULL) {
 884                 DEBUG(DEBUG_ERR, ("Failed to create a new lock context\n"));
 885                 return NULL;
 886         }
 887
 888         if ((request = talloc_zero(mem_ctx, struct lock_request)) == NULL) {
 889                 talloc_free(lock_ctx);
 890                 return NULL;
 891         }
 892
 893         lock_ctx->type = type;
 894         lock_ctx->ctdb = ctdb;
 895         lock_ctx->ctdb_db = ctdb_db;
 896         lock_ctx->key.dsize = key.dsize;
 897         if (key.dsize > 0) {
 898                 lock_ctx->key.dptr = talloc_memdup(lock_ctx, key.dptr, key.dsize);
 899                 if (lock_ctx->key.dptr == NULL) {
 900                         DEBUG(DEBUG_ERR, (__location__ "Memory allocation error\n"));
 901                         talloc_free(lock_ctx);
 902                         return NULL;
 903                 }
 904                 lock_ctx->key_hash = ctdb_hash(&key);
 905         } else {
 906                 lock_ctx->key.dptr = NULL;
 907         }
 908         lock_ctx->priority = priority;
 909         lock_ctx->auto_mark = auto_mark;
 910
 911         lock_ctx->request = request;
 912         lock_ctx->child = -1;
 913
 914         /* Non-record locks are required by recovery and should be scheduled
 915          * immediately, so keep them at the head of the pending queue.
 916          */
 917         if (lock_ctx->type == LOCK_RECORD) {
 918                 DLIST_ADD_END(ctdb_db->lock_pending, lock_ctx, NULL);
 919         } else {
 920                 DLIST_ADD_END(ctdb->lock_pending, lock_ctx, NULL);
 921         }
 922         CTDB_INCREMENT_STAT(ctdb, locks.num_pending);
 923         if (ctdb_db) {
 924                 CTDB_INCREMENT_DB_STAT(ctdb_db, locks.num_pending);
 925         }
 926
 927         /* Start the timer when we activate the context */
 928         lock_ctx->start_time = timeval_current();
 929
 930         request->lctx = lock_ctx;
 931         request->callback = callback;
 932         request->private_data = private_data;
 933
 934         talloc_set_destructor(request, ctdb_lock_request_destructor);
 935
 936         ctdb_lock_schedule(ctdb);
 937
 938         return request;
 939 }
 940
 941
 942 /*
 943  * obtain a lock on a record in a database
 944  */
 945 struct lock_request *ctdb_lock_record(TALLOC_CTX *mem_ctx,
 946                                       struct ctdb_db_context *ctdb_db,
 947                                       TDB_DATA key,
 948                                       bool auto_mark,
 949                                       void (*callback)(void *, bool),
 950                                       void *private_data)
 951 {
 952         return ctdb_lock_internal(mem_ctx,
 953                                   ctdb_db->ctdb,
 954                                   ctdb_db,
 955                                   key,
 956                                   0,
 957                                   callback,
 958                                   private_data,
 959                                   LOCK_RECORD,
 960                                   auto_mark);
 961 }
 962
 963
 964 /*
 965  * obtain a lock on a database
 966  */
 967 struct lock_request *ctdb_lock_db(TALLOC_CTX *mem_ctx,
 968                                   struct ctdb_db_context *ctdb_db,
 969                                   bool auto_mark,
 970                                   void (*callback)(void *, bool),
 971                                   void *private_data)
 972 {
 973         return ctdb_lock_internal(mem_ctx,
 974                                   ctdb_db->ctdb,
 975                                   ctdb_db,
 976                                   tdb_null,
 977                                   0,
 978                                   callback,
 979                                   private_data,
 980                                   LOCK_DB,
 981                                   auto_mark);
 982 }
 983
 984
 985 /*
 986  * obtain locks on all databases of specified priority
 987  */
 988 struct lock_request *ctdb_lock_alldb_prio(TALLOC_CTX *mem_ctx,
 989                                           struct ctdb_context *ctdb,
 990                                           uint32_t priority,
 991                                           bool auto_mark,
 992                                           void (*callback)(void *, bool),
 993                                           void *private_data)
 994 {
 995         if (priority < 1 || priority > NUM_DB_PRIORITIES) {
 996                 DEBUG(DEBUG_ERR, ("Invalid db priority: %u\n", priority));
 997                 return NULL;
 998         }
 999
1000         return ctdb_lock_internal(mem_ctx,
1001                                   ctdb,
1002                                   NULL,
1003                                   tdb_null,
1004                                   priority,
1005                                   callback,
1006                                   private_data,
1007                                   LOCK_ALLDB_PRIO,
1008                                   auto_mark);
1009 }
1010
1011
1012 /*
1013  * obtain locks on all databases
1014  */
1015 struct lock_request *ctdb_lock_alldb(TALLOC_CTX *mem_ctx,
1016                                      struct ctdb_context *ctdb,
1017                                      bool auto_mark,
1018                                      void (*callback)(void *, bool),
1019                                      void *private_data)
1020 {
1021         return ctdb_lock_internal(mem_ctx,
1022                                   ctdb,
1023                                   NULL,
1024                                   tdb_null,
1025                                   0,
1026                                   callback,
1027                                   private_data,
1028                                   LOCK_ALLDB,
1029                                   auto_mark);
1030 }
1031